Development and explanation of a machine learning model for identifying non-localized early-onset T1 colorectal cancer.

[BACKGROUND] For patients with early-onset T1 colorectal cancer (CRC), it is crucial to perform radical resection with minimal tissue damage. This study aims to develop and elucidate a machine learning model for the identification of non-located early-onset T1 CRC.

[METHODS] We extracted relevant data from the Surveillance, Epidemiology, and End Results (SEER) database and randomly allocated patients into a training set and a validation set. Five machine learning models were utilized: extreme gradient boosting (XGBoost), support vector machine (SVM), random forest (RF), decision tree (DT), and logistic regression (LR) were developed for the tumor metastasis status prediction. Area under the receiver operating characteristic curve (AUC) and confusion matrix were used to evaluate the model. The Shapley Additive Explanations (SHAP) method was employed as the primary technique for model explanation.

[RESULTS] A cohort of 1,878 patients was selected according to predefined inclusion and exclusion criteria. The XGBoost model was selected for its favorable performance, with an AUC of 0.755 and accuracy of 0.833 in the validation cohort. Significant risk factors, identified by SHAP values greater than 0.1 for non-localized T1 tumors, included carcinoembryonic antigen (CEA) levels exceeding 5 ng/mL, tumor size greater than 5 cm, mucinous adenocarcinoma, signet-ring cell adenocarcinoma, tumor location in the left colon, and poorly differentiated or undifferentiated histological grades. Additionally, black race, a diagnostic-to-surgical interval exceeding one month and PNI may represent potential risk factors.

[CONCLUSION] The XGBoost model holds the potential to assist clinicians in evaluating and selecting optimal treatment strategies for patients with early-onset T1 CRC.

[SUPPLEMENTARY INFORMATION] The online version contains supplementary material available at 10.1007/s12672-025-03998-8.